Merge branch 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[sfrench/cifs-2.6.git] / drivers / gpu / drm / i915 / intel_runtime_pm.c
1 /*
2  * Copyright © 2012-2014 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  * Authors:
24  *    Eugeni Dodonov <eugeni.dodonov@intel.com>
25  *    Daniel Vetter <daniel.vetter@ffwll.ch>
26  *
27  */
28
29 #include <linux/pm_runtime.h>
30 #include <linux/vgaarb.h>
31
32 #include <drm/drm_print.h>
33
34 #include "i915_drv.h"
35
36 /**
37  * DOC: runtime pm
38  *
39  * The i915 driver supports dynamic enabling and disabling of entire hardware
40  * blocks at runtime. This is especially important on the display side where
41  * software is supposed to control many power gates manually on recent hardware,
42  * since on the GT side a lot of the power management is done by the hardware.
43  * But even there some manual control at the device level is required.
44  *
45  * Since i915 supports a diverse set of platforms with a unified codebase and
46  * hardware engineers just love to shuffle functionality around between power
47  * domains there's a sizeable amount of indirection required. This file provides
48  * generic functions to the driver for grabbing and releasing references for
49  * abstract power domains. It then maps those to the actual power wells
50  * present for a given platform.
51  */
52
53 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_RUNTIME_PM)
54
55 #include <linux/sort.h>
56
57 #define STACKDEPTH 8
58
59 static noinline depot_stack_handle_t __save_depot_stack(void)
60 {
61         unsigned long entries[STACKDEPTH];
62         unsigned int n;
63
64         n = stack_trace_save(entries, ARRAY_SIZE(entries), 1);
65         return stack_depot_save(entries, n, GFP_NOWAIT | __GFP_NOWARN);
66 }
67
68 static void __print_depot_stack(depot_stack_handle_t stack,
69                                 char *buf, int sz, int indent)
70 {
71         unsigned long *entries;
72         unsigned int nr_entries;
73
74         nr_entries = stack_depot_fetch(stack, &entries);
75         stack_trace_snprint(buf, sz, entries, nr_entries, indent);
76 }
77
78 static void init_intel_runtime_pm_wakeref(struct intel_runtime_pm *rpm)
79 {
80         spin_lock_init(&rpm->debug.lock);
81 }
82
83 static noinline depot_stack_handle_t
84 track_intel_runtime_pm_wakeref(struct intel_runtime_pm *rpm)
85 {
86         depot_stack_handle_t stack, *stacks;
87         unsigned long flags;
88
89         if (!rpm->available)
90                 return -1;
91
92         stack = __save_depot_stack();
93         if (!stack)
94                 return -1;
95
96         spin_lock_irqsave(&rpm->debug.lock, flags);
97
98         if (!rpm->debug.count)
99                 rpm->debug.last_acquire = stack;
100
101         stacks = krealloc(rpm->debug.owners,
102                           (rpm->debug.count + 1) * sizeof(*stacks),
103                           GFP_NOWAIT | __GFP_NOWARN);
104         if (stacks) {
105                 stacks[rpm->debug.count++] = stack;
106                 rpm->debug.owners = stacks;
107         } else {
108                 stack = -1;
109         }
110
111         spin_unlock_irqrestore(&rpm->debug.lock, flags);
112
113         return stack;
114 }
115
116 static void untrack_intel_runtime_pm_wakeref(struct intel_runtime_pm *rpm,
117                                              depot_stack_handle_t stack)
118 {
119         unsigned long flags, n;
120         bool found = false;
121
122         if (unlikely(stack == -1))
123                 return;
124
125         spin_lock_irqsave(&rpm->debug.lock, flags);
126         for (n = rpm->debug.count; n--; ) {
127                 if (rpm->debug.owners[n] == stack) {
128                         memmove(rpm->debug.owners + n,
129                                 rpm->debug.owners + n + 1,
130                                 (--rpm->debug.count - n) * sizeof(stack));
131                         found = true;
132                         break;
133                 }
134         }
135         spin_unlock_irqrestore(&rpm->debug.lock, flags);
136
137         if (WARN(!found,
138                  "Unmatched wakeref (tracking %lu), count %u\n",
139                  rpm->debug.count, atomic_read(&rpm->wakeref_count))) {
140                 char *buf;
141
142                 buf = kmalloc(PAGE_SIZE, GFP_NOWAIT | __GFP_NOWARN);
143                 if (!buf)
144                         return;
145
146                 __print_depot_stack(stack, buf, PAGE_SIZE, 2);
147                 DRM_DEBUG_DRIVER("wakeref %x from\n%s", stack, buf);
148
149                 stack = READ_ONCE(rpm->debug.last_release);
150                 if (stack) {
151                         __print_depot_stack(stack, buf, PAGE_SIZE, 2);
152                         DRM_DEBUG_DRIVER("wakeref last released at\n%s", buf);
153                 }
154
155                 kfree(buf);
156         }
157 }
158
159 static int cmphandle(const void *_a, const void *_b)
160 {
161         const depot_stack_handle_t * const a = _a, * const b = _b;
162
163         if (*a < *b)
164                 return -1;
165         else if (*a > *b)
166                 return 1;
167         else
168                 return 0;
169 }
170
171 static void
172 __print_intel_runtime_pm_wakeref(struct drm_printer *p,
173                                  const struct intel_runtime_pm_debug *dbg)
174 {
175         unsigned long i;
176         char *buf;
177
178         buf = kmalloc(PAGE_SIZE, GFP_NOWAIT | __GFP_NOWARN);
179         if (!buf)
180                 return;
181
182         if (dbg->last_acquire) {
183                 __print_depot_stack(dbg->last_acquire, buf, PAGE_SIZE, 2);
184                 drm_printf(p, "Wakeref last acquired:\n%s", buf);
185         }
186
187         if (dbg->last_release) {
188                 __print_depot_stack(dbg->last_release, buf, PAGE_SIZE, 2);
189                 drm_printf(p, "Wakeref last released:\n%s", buf);
190         }
191
192         drm_printf(p, "Wakeref count: %lu\n", dbg->count);
193
194         sort(dbg->owners, dbg->count, sizeof(*dbg->owners), cmphandle, NULL);
195
196         for (i = 0; i < dbg->count; i++) {
197                 depot_stack_handle_t stack = dbg->owners[i];
198                 unsigned long rep;
199
200                 rep = 1;
201                 while (i + 1 < dbg->count && dbg->owners[i + 1] == stack)
202                         rep++, i++;
203                 __print_depot_stack(stack, buf, PAGE_SIZE, 2);
204                 drm_printf(p, "Wakeref x%lu taken at:\n%s", rep, buf);
205         }
206
207         kfree(buf);
208 }
209
210 static noinline void
211 __untrack_all_wakerefs(struct intel_runtime_pm_debug *debug,
212                        struct intel_runtime_pm_debug *saved)
213 {
214         *saved = *debug;
215
216         debug->owners = NULL;
217         debug->count = 0;
218         debug->last_release = __save_depot_stack();
219 }
220
221 static void
222 dump_and_free_wakeref_tracking(struct intel_runtime_pm_debug *debug)
223 {
224         struct drm_printer p;
225
226         if (!debug->count)
227                 return;
228
229         p = drm_debug_printer("i915");
230         __print_intel_runtime_pm_wakeref(&p, debug);
231
232         kfree(debug->owners);
233 }
234
235 static noinline void
236 __intel_wakeref_dec_and_check_tracking(struct intel_runtime_pm *rpm)
237 {
238         struct intel_runtime_pm_debug dbg = {};
239         unsigned long flags;
240
241         if (!atomic_dec_and_lock_irqsave(&rpm->wakeref_count,
242                                          &rpm->debug.lock,
243                                          flags))
244                 return;
245
246         __untrack_all_wakerefs(&rpm->debug, &dbg);
247         spin_unlock_irqrestore(&rpm->debug.lock, flags);
248
249         dump_and_free_wakeref_tracking(&dbg);
250 }
251
252 static noinline void
253 untrack_all_intel_runtime_pm_wakerefs(struct intel_runtime_pm *rpm)
254 {
255         struct intel_runtime_pm_debug dbg = {};
256         unsigned long flags;
257
258         spin_lock_irqsave(&rpm->debug.lock, flags);
259         __untrack_all_wakerefs(&rpm->debug, &dbg);
260         spin_unlock_irqrestore(&rpm->debug.lock, flags);
261
262         dump_and_free_wakeref_tracking(&dbg);
263 }
264
265 void print_intel_runtime_pm_wakeref(struct intel_runtime_pm *rpm,
266                                     struct drm_printer *p)
267 {
268         struct intel_runtime_pm_debug dbg = {};
269
270         do {
271                 unsigned long alloc = dbg.count;
272                 depot_stack_handle_t *s;
273
274                 spin_lock_irq(&rpm->debug.lock);
275                 dbg.count = rpm->debug.count;
276                 if (dbg.count <= alloc) {
277                         memcpy(dbg.owners,
278                                rpm->debug.owners,
279                                dbg.count * sizeof(*s));
280                 }
281                 dbg.last_acquire = rpm->debug.last_acquire;
282                 dbg.last_release = rpm->debug.last_release;
283                 spin_unlock_irq(&rpm->debug.lock);
284                 if (dbg.count <= alloc)
285                         break;
286
287                 s = krealloc(dbg.owners,
288                              dbg.count * sizeof(*s),
289                              GFP_NOWAIT | __GFP_NOWARN);
290                 if (!s)
291                         goto out;
292
293                 dbg.owners = s;
294         } while (1);
295
296         __print_intel_runtime_pm_wakeref(p, &dbg);
297
298 out:
299         kfree(dbg.owners);
300 }
301
302 #else
303
304 static void init_intel_runtime_pm_wakeref(struct intel_runtime_pm *rpm)
305 {
306 }
307
308 static depot_stack_handle_t
309 track_intel_runtime_pm_wakeref(struct intel_runtime_pm *rpm)
310 {
311         return -1;
312 }
313
314 static void untrack_intel_runtime_pm_wakeref(struct intel_runtime_pm *rpm,
315                                              intel_wakeref_t wref)
316 {
317 }
318
319 static void
320 __intel_wakeref_dec_and_check_tracking(struct intel_runtime_pm *rpm)
321 {
322         atomic_dec(&rpm->wakeref_count);
323 }
324
325 static void
326 untrack_all_intel_runtime_pm_wakerefs(struct intel_runtime_pm *rpm)
327 {
328 }
329
330 #endif
331
332 static void
333 intel_runtime_pm_acquire(struct intel_runtime_pm *rpm, bool wakelock)
334 {
335         if (wakelock) {
336                 atomic_add(1 + INTEL_RPM_WAKELOCK_BIAS, &rpm->wakeref_count);
337                 assert_rpm_wakelock_held(rpm);
338         } else {
339                 atomic_inc(&rpm->wakeref_count);
340                 assert_rpm_raw_wakeref_held(rpm);
341         }
342 }
343
344 static void
345 intel_runtime_pm_release(struct intel_runtime_pm *rpm, int wakelock)
346 {
347         if (wakelock) {
348                 assert_rpm_wakelock_held(rpm);
349                 atomic_sub(INTEL_RPM_WAKELOCK_BIAS, &rpm->wakeref_count);
350         } else {
351                 assert_rpm_raw_wakeref_held(rpm);
352         }
353
354         __intel_wakeref_dec_and_check_tracking(rpm);
355 }
356
357 static intel_wakeref_t __intel_runtime_pm_get(struct intel_runtime_pm *rpm,
358                                               bool wakelock)
359 {
360         int ret;
361
362         ret = pm_runtime_get_sync(rpm->kdev);
363         WARN_ONCE(ret < 0, "pm_runtime_get_sync() failed: %d\n", ret);
364
365         intel_runtime_pm_acquire(rpm, wakelock);
366
367         return track_intel_runtime_pm_wakeref(rpm);
368 }
369
370 /**
371  * intel_runtime_pm_get_raw - grab a raw runtime pm reference
372  * @rpm: the intel_runtime_pm structure
373  *
374  * This is the unlocked version of intel_display_power_is_enabled() and should
375  * only be used from error capture and recovery code where deadlocks are
376  * possible.
377  * This function grabs a device-level runtime pm reference (mostly used for
378  * asynchronous PM management from display code) and ensures that it is powered
379  * up. Raw references are not considered during wakelock assert checks.
380  *
381  * Any runtime pm reference obtained by this function must have a symmetric
382  * call to intel_runtime_pm_put_raw() to release the reference again.
383  *
384  * Returns: the wakeref cookie to pass to intel_runtime_pm_put_raw(), evaluates
385  * as True if the wakeref was acquired, or False otherwise.
386  */
387 intel_wakeref_t intel_runtime_pm_get_raw(struct intel_runtime_pm *rpm)
388 {
389         return __intel_runtime_pm_get(rpm, false);
390 }
391
392 /**
393  * intel_runtime_pm_get - grab a runtime pm reference
394  * @rpm: the intel_runtime_pm structure
395  *
396  * This function grabs a device-level runtime pm reference (mostly used for GEM
397  * code to ensure the GTT or GT is on) and ensures that it is powered up.
398  *
399  * Any runtime pm reference obtained by this function must have a symmetric
400  * call to intel_runtime_pm_put() to release the reference again.
401  *
402  * Returns: the wakeref cookie to pass to intel_runtime_pm_put()
403  */
404 intel_wakeref_t intel_runtime_pm_get(struct intel_runtime_pm *rpm)
405 {
406         return __intel_runtime_pm_get(rpm, true);
407 }
408
409 /**
410  * intel_runtime_pm_get_if_in_use - grab a runtime pm reference if device in use
411  * @rpm: the intel_runtime_pm structure
412  *
413  * This function grabs a device-level runtime pm reference if the device is
414  * already in use and ensures that it is powered up. It is illegal to try
415  * and access the HW should intel_runtime_pm_get_if_in_use() report failure.
416  *
417  * Any runtime pm reference obtained by this function must have a symmetric
418  * call to intel_runtime_pm_put() to release the reference again.
419  *
420  * Returns: the wakeref cookie to pass to intel_runtime_pm_put(), evaluates
421  * as True if the wakeref was acquired, or False otherwise.
422  */
423 intel_wakeref_t intel_runtime_pm_get_if_in_use(struct intel_runtime_pm *rpm)
424 {
425         if (IS_ENABLED(CONFIG_PM)) {
426                 /*
427                  * In cases runtime PM is disabled by the RPM core and we get
428                  * an -EINVAL return value we are not supposed to call this
429                  * function, since the power state is undefined. This applies
430                  * atm to the late/early system suspend/resume handlers.
431                  */
432                 if (pm_runtime_get_if_in_use(rpm->kdev) <= 0)
433                         return 0;
434         }
435
436         intel_runtime_pm_acquire(rpm, true);
437
438         return track_intel_runtime_pm_wakeref(rpm);
439 }
440
441 /**
442  * intel_runtime_pm_get_noresume - grab a runtime pm reference
443  * @rpm: the intel_runtime_pm structure
444  *
445  * This function grabs a device-level runtime pm reference (mostly used for GEM
446  * code to ensure the GTT or GT is on).
447  *
448  * It will _not_ power up the device but instead only check that it's powered
449  * on.  Therefore it is only valid to call this functions from contexts where
450  * the device is known to be powered up and where trying to power it up would
451  * result in hilarity and deadlocks. That pretty much means only the system
452  * suspend/resume code where this is used to grab runtime pm references for
453  * delayed setup down in work items.
454  *
455  * Any runtime pm reference obtained by this function must have a symmetric
456  * call to intel_runtime_pm_put() to release the reference again.
457  *
458  * Returns: the wakeref cookie to pass to intel_runtime_pm_put()
459  */
460 intel_wakeref_t intel_runtime_pm_get_noresume(struct intel_runtime_pm *rpm)
461 {
462         assert_rpm_wakelock_held(rpm);
463         pm_runtime_get_noresume(rpm->kdev);
464
465         intel_runtime_pm_acquire(rpm, true);
466
467         return track_intel_runtime_pm_wakeref(rpm);
468 }
469
470 static void __intel_runtime_pm_put(struct intel_runtime_pm *rpm,
471                                    intel_wakeref_t wref,
472                                    bool wakelock)
473 {
474         struct device *kdev = rpm->kdev;
475
476         untrack_intel_runtime_pm_wakeref(rpm, wref);
477
478         intel_runtime_pm_release(rpm, wakelock);
479
480         pm_runtime_mark_last_busy(kdev);
481         pm_runtime_put_autosuspend(kdev);
482 }
483
484 /**
485  * intel_runtime_pm_put_raw - release a raw runtime pm reference
486  * @rpm: the intel_runtime_pm structure
487  * @wref: wakeref acquired for the reference that is being released
488  *
489  * This function drops the device-level runtime pm reference obtained by
490  * intel_runtime_pm_get_raw() and might power down the corresponding
491  * hardware block right away if this is the last reference.
492  */
493 void
494 intel_runtime_pm_put_raw(struct intel_runtime_pm *rpm, intel_wakeref_t wref)
495 {
496         __intel_runtime_pm_put(rpm, wref, false);
497 }
498
499 /**
500  * intel_runtime_pm_put_unchecked - release an unchecked runtime pm reference
501  * @rpm: the intel_runtime_pm structure
502  *
503  * This function drops the device-level runtime pm reference obtained by
504  * intel_runtime_pm_get() and might power down the corresponding
505  * hardware block right away if this is the last reference.
506  *
507  * This function exists only for historical reasons and should be avoided in
508  * new code, as the correctness of its use cannot be checked. Always use
509  * intel_runtime_pm_put() instead.
510  */
511 void intel_runtime_pm_put_unchecked(struct intel_runtime_pm *rpm)
512 {
513         __intel_runtime_pm_put(rpm, -1, true);
514 }
515
516 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_RUNTIME_PM)
517 /**
518  * intel_runtime_pm_put - release a runtime pm reference
519  * @rpm: the intel_runtime_pm structure
520  * @wref: wakeref acquired for the reference that is being released
521  *
522  * This function drops the device-level runtime pm reference obtained by
523  * intel_runtime_pm_get() and might power down the corresponding
524  * hardware block right away if this is the last reference.
525  */
526 void intel_runtime_pm_put(struct intel_runtime_pm *rpm, intel_wakeref_t wref)
527 {
528         __intel_runtime_pm_put(rpm, wref, true);
529 }
530 #endif
531
532 /**
533  * intel_runtime_pm_enable - enable runtime pm
534  * @rpm: the intel_runtime_pm structure
535  *
536  * This function enables runtime pm at the end of the driver load sequence.
537  *
538  * Note that this function does currently not enable runtime pm for the
539  * subordinate display power domains. That is done by
540  * intel_power_domains_enable().
541  */
542 void intel_runtime_pm_enable(struct intel_runtime_pm *rpm)
543 {
544         struct device *kdev = rpm->kdev;
545
546         /*
547          * Disable the system suspend direct complete optimization, which can
548          * leave the device suspended skipping the driver's suspend handlers
549          * if the device was already runtime suspended. This is needed due to
550          * the difference in our runtime and system suspend sequence and
551          * becaue the HDA driver may require us to enable the audio power
552          * domain during system suspend.
553          */
554         dev_pm_set_driver_flags(kdev, DPM_FLAG_NEVER_SKIP);
555
556         pm_runtime_set_autosuspend_delay(kdev, 10000); /* 10s */
557         pm_runtime_mark_last_busy(kdev);
558
559         /*
560          * Take a permanent reference to disable the RPM functionality and drop
561          * it only when unloading the driver. Use the low level get/put helpers,
562          * so the driver's own RPM reference tracking asserts also work on
563          * platforms without RPM support.
564          */
565         if (!rpm->available) {
566                 int ret;
567
568                 pm_runtime_dont_use_autosuspend(kdev);
569                 ret = pm_runtime_get_sync(kdev);
570                 WARN(ret < 0, "pm_runtime_get_sync() failed: %d\n", ret);
571         } else {
572                 pm_runtime_use_autosuspend(kdev);
573         }
574
575         /*
576          * The core calls the driver load handler with an RPM reference held.
577          * We drop that here and will reacquire it during unloading in
578          * intel_power_domains_fini().
579          */
580         pm_runtime_put_autosuspend(kdev);
581 }
582
583 void intel_runtime_pm_disable(struct intel_runtime_pm *rpm)
584 {
585         struct device *kdev = rpm->kdev;
586
587         /* Transfer rpm ownership back to core */
588         WARN(pm_runtime_get_sync(kdev) < 0,
589              "Failed to pass rpm ownership back to core\n");
590
591         pm_runtime_dont_use_autosuspend(kdev);
592
593         if (!rpm->available)
594                 pm_runtime_put(kdev);
595 }
596
597 void intel_runtime_pm_cleanup(struct intel_runtime_pm *rpm)
598 {
599         int count = atomic_read(&rpm->wakeref_count);
600
601         WARN(count,
602              "i915 raw-wakerefs=%d wakelocks=%d on cleanup\n",
603              intel_rpm_raw_wakeref_count(count),
604              intel_rpm_wakelock_count(count));
605
606         untrack_all_intel_runtime_pm_wakerefs(rpm);
607 }
608
609 void intel_runtime_pm_init_early(struct intel_runtime_pm *rpm)
610 {
611         struct drm_i915_private *i915 =
612                         container_of(rpm, struct drm_i915_private, runtime_pm);
613         struct pci_dev *pdev = i915->drm.pdev;
614         struct device *kdev = &pdev->dev;
615
616         rpm->kdev = kdev;
617         rpm->available = HAS_RUNTIME_PM(i915);
618
619         init_intel_runtime_pm_wakeref(rpm);
620 }